Decomposition and Its Logic
Decomposition is the operation by which bounded agents convert problems too complex to solve directly into collections of sub-problems each solvable by finite cognition. The practice is ubiquitous because bounded rationality requires it: no mind can hold an entire complex system in view, so complex systems must be built, maintained, and modified through decomposition into subsystems that bounded minds can understand separately. Simon's near-decomposability framework provides the structural conditions under which decomposition works — subsystems must interact strongly within themselves and weakly across boundaries, so that understanding any particular subsystem does not require understanding all of them simultaneously. The traditional corporate hierarchy, the academic department, the software module, the engineering specialty — all are decomposition architectures, and all are being destabilized by AI's capacity to enable individuals to operate across subsystem boundaries. The question for the next generation of organizations and institutions is not whether decomposition will survive but what new decompositions will replace the old ones — which boundaries will be drawn where, and what interfaces will carry the information that decomposition requires.
In the AI Story
The logic of decomposition is cognitive before it is organizational. A bounded mind cannot hold a thousand-component system in working memory; it can hold a ten-component sub-assembly. Decomposition respects this limit by organizing work into sub-assembly-sized chunks, each small enough for bounded cognition and connected to other chunks through narrow, well-defined interfaces. The resulting structure is buildable by finite minds in a way that the undecomposed system is not.
Every major organizational innovation in the history of modern industry has been a decomposition innovation — the assembly line decomposed manufacturing by task, the corporate hierarchy decomposed management by function, the academic discipline decomposed inquiry by domain. Each decomposition enabled bounded minds to coordinate on work that exceeded any individual's capacity, and each one embedded assumptions about what the bounded minds were bounded by. When the bounds shift — when AI enables individuals to operate across domains that the decomposition had separated — the decomposition itself becomes obsolete.
The AI age is producing a period of decomposition experimentation that will likely last for a generation. The old decompositions are being dismantled. New decompositions — vector pods that organize work by cognitive operation rather than functional domain, human-AI teams that separate generation from evaluation, judgment-intensive roles that specialize in the evaluative work AI cannot perform — are being invented in real time. Some will stabilize into durable structures; most will not. The question of which decompositions serve the new cognitive environment well, and which merely reproduce the old patterns in new language, will take decades to settle.
Origin
The logic of decomposition appears throughout Simon's work, from Administrative Behavior (1947) through The Sciences of the Artificial (1969) and beyond. The most systematic statement is in 'The Architecture of Complexity' (1962), which articulated near-decomposability as the universal structural principle making decomposition viable.
The AI-era implications of decomposition theory were not explicit in Simon's original formulation — he did not live to see large language models — but follow directly from his framework. If decomposition exists because bounded minds require it, then changes to what bounded minds can manage will change what decompositions are required. AI has changed what bounded minds can manage. New decompositions follow.
Key Ideas
Decomposition is cognitive. The practice exists because bounded minds cannot hold complex systems in view; it is the structural prerequisite for bounded agents building complex artifacts.
Near-decomposability is the structural requirement. Decomposition works when subsystems interact strongly within themselves and weakly across boundaries; it fails when the boundaries carry more traffic than the interfaces can accommodate.
Organizations are decompositions. Hierarchies, departments, specializations are all decomposition architectures, reflecting assumptions about what bounded minds can manage.
AI destabilizes old decompositions. When individuals can operate across subsystem boundaries, the boundaries lose their architectural function.
New decompositions are emerging. Vector pods, human-AI teams, judgment-focused roles represent the current experimental response to the shifted cognitive environment.
Appears in the Orange Pill Cycle
Further reading
- Simon, 'The Architecture of Complexity' (1962)
- Simon, Administrative Behavior (1947)
- Carliss Baldwin and Kim Clark, Design Rules (2000)
- Adam Smith, The Wealth of Nations (1776, Book I)